stable-dreamfusion
ashawkey/stable-dreamfusion · Python · Apache-2.0
Classic open text-to-3D playground
stable-dreamfusion popularized accessible DreamFusion-style experiments. Newer feed-forward methods are faster, but this repo remains educational.
Score
Setup at a glance
git clone https://github.com/ashawkey/stable-dreamfusion.git Deep dive
stable-dreamfusion is the open implementation of DreamFusion-style text-to-3D using Stable Diffusion: no 3D training data needed — the 2D diffusion model itself supervises the NeRF optimization.
It supports Perp-Neg (perpendicular negative prompting) to alleviate the multi-face Janus problem, offers both the original DeepFloyd IF and Stable Diffusion backbones, and an Instant-NGP backbone (-O) for much faster optimization.
It remains a teaching classic: the cleanest way to understand score distillation sampling and the trade-offs of text-to-3D before using modern feed-forward models.